AI Image & Video Model News Roundup – October 24, 2026

A week of follow-through: the models from last week’s releases are already getting editors, distillations, and upscalers, Tencent shipped a world model you can steer live, and ComfyUI pushed another core release. Here are the highlights from October 18 to October 24, 2026.

New image models

Fibo 2.0: one 8B model that generates and edits images

Bria released Fibo 2.0 on October 18-19: an 8B flow-matching model that handles both image generation and editing in a single set of weights. Two checkpoints ship: Dev, the full 50-step base meant for fine-tuning and research, and Edit Turbo, a 4-step distilled checkpoint for low-latency edits. Editing covers instruction edits, masked edits, and multi-reference edits with up to 4 images, and the model card cites an Artificial Analysis snapshot ranking it first among open-weights models on image editing. The catch for tinkerers: prompts are structured JSON documents (Bria’s VGL format) rather than plain sentences, though a helper package turns plain instructions into that format. Native ComfyUI support landed in core on October 21, with Comfy-Org repacks in BF16 and INT8. Details at the ComfyUI Wiki. Note the weights are gated on Hugging Face, so you accept Bria’s terms before downloading.

MOSS-Image gets its first instruction-editing LoRA

Last week’s layout model can now edit, too. On October 22 the community released a rank-64 MOSS-Image Edit LoRA that takes one to four reference images plus a text instruction and changes color, outfits, lighting, seasons, and style while the references hold the rest steady. With multiple references you name each image’s role in the instruction, for example “dress the person in image 1 in the outfit from image 2.” Remarkably, ComfyUI core merged native support for exactly this adapter class within nine hours, via a new experimental Text Encode MOSS-Image Edit node, plus a speed fix that nearly halves per-step time for 1024-pixel edits with a 1 MP reference. One caveat: the core support merged after the v0.42.1 tag, so you need a build newer than that release to use it. Full breakdown at the ComfyUI Wiki.

MOSS-Image MAX ULTRA LoHa pushes output to 4 megapixels and beyond

Also in MOSS-Image land: comfyanonymous published a LoHa that trains MOSS-Image-Base for 4 MP and larger canvases, shipping four checkpoint versions with 4096×2304 comparison samples. If last week’s 50-step MOSS workflow felt resolution-limited, this is the adapter to try. See the ComfyUI Wiki news page for the write-up.

New video models

Hy-Nomas: Tencent’s interactive world model you steer with text and camera moves

Tencent Hunyuan released Hy-Nomas on October 21: an interactive world model where you describe a world, then keep steering it with text instructions and camera trajectories while it keeps generating. It is autoregressive across video segments, runs a distilled 8-step configuration at 720p and 24 FPS, and ships with weights, inference code, training code, and a benchmark of 280 multi-turn test cases. There is even a single-RTX-4090 configuration using INT8 weights and an FP8 decoder, though the team is upfront that speed there is unverified. A community ComfyUI port arrived October 23 with converted INT8 and BF16 weights that load through stock nodes. The Wiki breakdown covers the camera conditioning and the 4090 trade-offs.

MiniMax H3 Ref2VA drops to 8 or 4 steps with PTD LoRAs

The few-step race for MiniMax H3 continues. Taobao Content AI released per-hop PTD LoRAs (Parametric Trajectory Distillation, from EPFL, NVIDIA, and Stanford researchers) that take H3 Ref2VA, reference images in and synced audio-video out, down to 8 or 4 transformer evaluations, with a RefAsymOPD sibling claiming better background plausibility and sharpness at the same step count. A community compact loader already makes them work on pruned and quantized bases, since the official route wants the unpruned 140 GB base. In the same week, FastVideo’s FastH3 Omni shipped an 8-step H3 distillation that turns image, video, and audio references into one clip with stereo sound. Details on PTD at the ComfyUI Wiki.

ComfyUI workflows and tools

ComfyUI v0.42.0 and v0.42.1: native MOSS-Image, Krea 3 partner node, MOSS template crash fix

Two more core releases. v0.42.0 (release notes) adds native MOSS-Image support, a Krea 3 Image partner node, Qwen 3.x performance work, and updated workflow templates. v0.42.1 (release notes) is a focused fix for crashes some users hit running the MOSS image template workflow. As usual, back up your install before upgrading your production box, and remember the MOSS-Image Edit node support above landed after v0.42.1, so track master if you want it now. The official changelog tracks the daily node updates, including new VEED and Mirelo partner nodes this week.

ComfyUI-RMBG v3.3.0: Canvas Studio and video matting

The background-removal staple got a big update. ComfyUI-RMBG v3.3.0 adds a Canvas compositing studio, VRMBG video matting, FeyNoBg fine-hair matting, YOLO26 segmentation, and an offline Florence-2 rebuild. If you cut subjects out of images or video in ComfyUI, this is the update to grab. See the ComfyUI Wiki news page for details.

Bottom line

This week the ecosystem did what it does best: it took last week’s models and made them more useful. MOSS-Image went from a layout generator to an editor and a 4 MP canvas in seven days, H3 video keeps collapsing toward real time, and Tencent’s Hy-Nomas points at a future where video models are worlds you walk through rather than clips you render. One deadline to keep on your radar: Comfy Org’s Open Source Higgsfield challenge closes October 30 at 9am PT, with $25,000 in prizes for open rebuilds of subscription creative apps. See you next Saturday.

Further Reading

https://www.kombitz.com/2026/10/18/how-to-use-flux-3-klein-gguf-workflow-in-comfyui/ https://www.kombitz.com/2026/10/18/ai-image-video-model-news-roundup-october-18-2026/ https://www.kombitz.com/2026/09/27/minimax-h3-video-vae-gets-up-to-2-2x-faster-in-comfyui/